AI agents solve 49 of 50 physics problems in DeepMind-backed study
AI Agents Solve 49 of 50 Unpublished Physics Problems, and Trading Bots Should Be Paying Attention
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
A group of physicists working with Google DeepMind backing set 50 unpublished theoretical physics problems in front of a set of AI agents, and the agents solved 49 of them (Crypto Briefing, 2026). That is a 98 percent solve rate on problems the models had never seen, which is the part that matters. Benchmarks a model has already trained on tell you almost nothing. Unpublished problems are the closest thing AI research has to a clean, out-of-sample test.
We spend our working lives inside a related but far messier question. We run six-month funded-account trials of AI trading bots, and the sub-niche we are evaluating here is the AI trading bot category specifically: systems that claim to reason about markets and place orders without a human clicking the button. When we read a 98 percent headline, our first instinct is not excitement. It is to ask what the answer key looks like. Physics has one. Markets do not. That single difference is where a lot of AI trading bot marketing quietly falls apart, and it is the thread we want to pull here.
We have benchmarked adaptive engines, including Zephyr AI's adaptive engine, against reasoning-heavy claims throughout our 2026 review cycle, and the pattern holds: the harder the problem, the more the marketing outruns the live profit and loss.
What did the DeepMind-backed study actually show?
The headline is simple. Fifty problems, forty-nine solved, drawn from unpublished theoretical physics work, with Google DeepMind backing the effort and physicists at ICTS involved. The summary framing is that near-perfect problem-solving in theoretical physics highlights AI's potential to improve research efficiency and accuracy (Crypto Briefing RSS summary, 2026).
What we do not have from the summary is the grading rubric, the time budget allowed per problem, or whether the agents had tool access, a calculator, or a search step. Those details decide whether 49 of 50 is a reasoning breakthrough or a well-engineered pipeline. We treat any single benchmark number with measured skepticism until the primary paper is public, and we would tell a retail trader the same thing about a bot vendor's glossy backtest page.
The genuinely interesting design choice is the word "unpublished." Problems the model has never seen are the only problems that tell you whether a system generalizes. That is the standard we try to hold AI trading bots to, and it is a standard most of them never meet.
Why a physics benchmark is not a trading benchmark
Here is the editorial point that the coverage largely skips. A physics problem has a ground truth. The answer is either right or wrong, and a physicist can check it. A trading decision has no such key. The only feedback loop is realized profit and loss, and that number is noisy, delayed, and generated by a market whose statistical properties keep shifting.
That asymmetry has three consequences for anyone evaluating an AI trading bot.
First, a 98 percent solve rate on a static benchmark does not transfer to a non-stationary environment. The physics problems do not change their answers between Monday and Friday. A market regime can invert a strategy's edge in a single session, and the agent gets no clean signal that it has happened.
Second, benchmark contamination is a live risk in trading that it is not in physics. When a vendor shows you a backtest, the vendor chose the period, the instrument, and the parameters. That is the opposite of an unpublished-problem test. It is closer to giving a student the exam paper a week early.
Third, and this is the under-discussed one, an AI trading agent can be right about the market and still lose money. A correct directional call placed with the wrong position size, at the wrong time, through a broker with wide spreads, produces a loss. The reasoning layer can score 98 percent while the execution and risk layers quietly destroy the account. In our funded-account trials, the failures we see most often are not reasoning failures. They are sizing and execution failures.
What does an AI trading bot actually do?
Strip away the branding and an AI trading bot has three jobs. It generates a signal, it decides how much to risk, and it sends an order to a broker or exchange. The reasoning breakthrough in the physics study maps onto the first job. The other two are where retail accounts live or die.
The category is broad. At one end you have rule-based expert advisors on MT4 and MT5, which are deterministic and easy to audit. In the middle sit machine-learning signal providers that publish a probability and leave execution to you. At the far end are agent-style systems that claim to reason about context and adapt in real time, which is the family the DeepMind-backed result makes more credible on paper.
Open-source frameworks such as Backtrader and NautilusTrader are commonly used as evaluation subjects for backtesting logic, and they are useful for that purpose. What they are not is a recommendation, because a framework is only as good as the strategy a human programs into it, and most retail traders do not have a strategy worth automating.
How big is the backtest-to-live gap, really?
The gap is always there and it is always real. It comes from slippage, latency, spread, regime change, and the simple fact that a backtest is fitted to the past. We do not have a published live-versus-backtest figure for the study's agent, because it is not a trading system, so we will not pretend otherwise. Any specific gap number you see for a bot should be verified directly with the provider, and you should ask whether the live window was chosen before or after the strategy was tuned.
What we can say from our testing framework is that we weight live-trade metrics far more heavily than any backtest, and we discount backtests that do not disclose their out-of-sample period. A vendor that shows only an equity curve and no live window has told you nothing you can act on.
| Dimension | Physics benchmark (source study) | AI trading bot evaluation |
|---|---|---|
| Ground truth | A correct answer exists for each problem | No answer key, only realized profit and loss |
| Sample | 50 unpublished problems | Varies by strategy, verify with provider |
| Headline result | 49 of 50 solved (98 percent) | Report live and backtest separately |
| Contamination risk | Low, problems were unpublished | High, the vendor chooses the backtest window |
| Failure mode | Wrong answer, visible immediately | Wrong position, visible only after losses accumulate |
How big are the drawdowns?
Drawdown is the number that decides whether a retail trader can actually stay in a strategy. A bot can post a respectable average return and still be unusable if a single bad week takes the account down 30 percent and the trader panics out at the bottom.
We do not have drawdown data for the study's agent, because drawdown is a trading concept and the study is not a trading study. For any specific bot, this is a verify-with-provider item: ask for maximum drawdown, the recovery period, and the worst rolling 30-day window. If a vendor cannot produce those three numbers, treat the strategy as unproven.
| Evaluation dimension | Data available in our review window | Where to verify |
|---|---|---|
| Strategy specification | Not published for the study agent | Primary paper and provider documentation |
| Live versus backtest gap | N/A for the study | Provider published metrics |
| Maximum drawdown | N/A for the study | Provider risk disclosure |
| Fee schedule | N/A for the study | Provider pricing page |
| Regulatory status | N/A, research is not a financial product | FCA Register, ASIC Connect |
| Withdrawal and disengagement | N/A | Provider terms of service |
Free Download: DeepMind-Backed AI Agent Trading Bot Due-Diligence Checklist: 49/50 Physics Problems to Live Markets
A due-diligence checklist to test whether an AI agent that solved 49 of 50 physics problems can also deliver transparent strategy specs, reliable backtests, broker compatibility, fees, and withdrawals.
Get the AI Agent Checklist
Not sure which AI trading bot fits your strategy? Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026
This link is an affiliate partnership - see our editorial policy for details.
What does it cost, and does the fee model eat the edge?
Subscription economics decide whether a bot is worth running at all. The common models are a flat monthly fee, a performance fee on profits, or a hybrid. The trap is the flat fee on a small account. If a bot charges a fixed monthly amount, that cost is a percentage of your capital that rises as your account shrinks, which is exactly when you can least afford it.
We do not have a fee schedule for the study, and we will not invent one. For a real bot, model the fee against your actual account size before you subscribe, not against the vendor's example account. A performance fee aligns interests better in theory, but read the definition of "profit" carefully, because some providers count gross gains before losses.
Can you actually stop the bot cleanly?
Disengagement is the least glamorous and most important test we run. When we end a trial, we check three things: whether open positions close cleanly, whether the API key can be revoked without leaving dangling orders, and whether the subscription cancels without a retention maze.
If a provider makes it hard to leave, treat that as a risk signal about how it treats customers generally. A clean exit should be a two-step process, not a support ticket and a phone call.
Is any of this regulated?
The physics study is research, not a financial product, so no financial regulator has any jurisdiction over it. That is fine, because nobody is being asked to deposit money with it.
AI trading bots are a different matter. Before we recommend any automated service, we check the vendor's claimed regulatory status against the primary register. In the United Kingdom that means the FCA Register, and in Australia it means the ASIC Connect registers. Neither register returns a listing tied to the study, which is expected, and any bot provider that claims a licence should be checked against the same registers before you fund an account. If the research data does not include the register entry, verify directly with the provider's primary regulator rather than taking a website badge at face value.
| Regulator | Register | What it confirms |
|---|---|---|
| FCA (United Kingdom) | FCA Register | Firm authorisation and permissions |
| ASIC (Australia) | ASIC Connect registers | AFSL and credit licence status |
| Other jurisdictions | Verify directly with the provider's primary regulator | Local licensing and conduct rules |
How Zephyr AI compares
The study's agent is a research artifact, so the fair comparison is not against a bot at all. It is against the transparency standard we apply to bots we actually test. On the concrete dimension of published fee structure and regulatory posture, Zephyr AI's platform publishes what it charges and how it is structured, while the study agent's deployment details are simply not public because it is not a commercial product.
Where Zephyr AI's adaptive position-sizing has an edge in our review cycle is in the layer the physics result does not touch: risk sizing under changing volatility. A benchmark that solves 49 of 50 problems says nothing about how a system sizes the 51st trade on a Friday afternoon with a central bank decision pending. That is the layer we care about, and it is the layer we will keep testing.
Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026
Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026
This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.
Frequently Asked Questions
What did the DeepMind-backed study actually prove?
It showed that AI agents could solve 49 of 50 unpublished theoretical physics problems, a 98 percent out-of-sample solve rate (Crypto Briefing, 2026). The key detail is that the problems were unpublished, so the agents had not seen them before, which makes the result a genuine generalization test rather than a memorization test.
Does a 98 percent solve rate mean an AI trading bot will be profitable?
No. Physics problems have a verifiable answer, while trading decisions only produce profit or loss in a shifting market with no answer key. A high benchmark score says nothing about position sizing, execution quality, or how a strategy behaves when the market regime changes.
Can I run an AI trading bot on a prop firm account?
It depends entirely on the prop firm's rules, and many firms ban fully automated execution or restrict it to specific account types. Check the firm's terms before you connect any bot, and confirm whether the provider's own terms permit prop-firm use.
Does this bot work in the US under Pattern Day Trader rules?
Pattern Day Trader rules apply to margin accounts with fewer than 25,000 dollars that place four or more day trades in five business days. Any high-frequency bot can trip that threshold quickly, so US-based traders should confirm their broker's treatment of automated day trades before running a strategy live.
What happens if the API connection drops mid-trade?
This is a critical failure point, and the answer varies by provider. A well-built bot should have a defined state-recovery process and a way to close or hedge orphaned positions, but you should verify the provider's documented behavior rather than assume it.
How do I verify an AI trading bot's drawdown claims?
Ask for the maximum drawdown, the recovery period, and the worst rolling 30-day window, and ask whether those figures come from live trading or a backtest. If the provider cannot separate live from simulated results, treat the numbers as unverified.
Is Zephyr AI regulated?
Regulatory status varies by jurisdiction and should be checked against the primary register. In the UK that is the FCA Register, and in Australia it is ASIC Connect. Always verify the specific entity and licence before funding an account.
Can I stop an AI trading bot cleanly and withdraw?
A clean exit should let you close open positions, revoke the API key, and cancel the subscription without a retention maze. Test the exit process on a small account before you scale up, because leaving a bot is often harder than starting one.
How big is the backtest-to-live gap for AI trading bots?
The gap is always present and it comes from slippage, latency, spread, and regime change. We do not have a universal figure, because it varies by strategy and market, so verify the provider's live window directly and discount any backtest that does not disclose its out-of-sample period.
Not sure which AI trading bot fits your strategy? Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026
This link is an affiliate partnership - see our editorial policy for details.
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.
More in this category: AI Trading Bot Reviews.